Resumo:
Predicting the Rate of Penetration (ROP) plays a central role in optimizing operations
in the oil and gas industry, directly influencing operational efficiency and drilling-related
costs. However, conventional machine learning models have limitations in this context due
to the strong non-linearity of the variables involved and, above all, the high lithological
heterogeneity and low reliability of downhole sensors, resulting in noisy and gapped time
series. This work proposes AutoDrill, an Automated Machine Learning (AutoML) framework
specifically designed for the drilling domain. In contrast to generalist approaches,
AutoDrill incorporates a preprocessing pipeline specialized in drilling signals (including
adaptive smoothing, transient correction, and data interpolation) coupled with an automated
hyperparameter search and tuning mechanism. The experimental evaluation was
conducted on various datasets and geological scenarios, using records from three offshore
repositories (Equinor Volve, USROP, and NEATS NOPIMS). The performance of AutoDrill
was compared to the AutoGluon and H2O AutoML frameworks, considered representative
of the state of the art in AutoML. The results indicate that AutoDrill exhibits consistent
and competitive performance, achieving an average coefficient of determination (R2) of
0.91 across all analyzed scenarios, particularly in environments characterized by high
heterogeneity and noise, such as in the case of NEATS NOPIMS. In this specific scenario,
AutoDrill significantly outperformed the reference frameworks, reducing the mean absolute
error (MAE) by up to 80% and raising the R2 from levels below 0.20 to 0.94. In conclusion,
the systematic incorporation of domain knowledge into the machine learning automation
cycle is crucial for the robustness of real-time prediction. From this perspective, AutoDrill
presents itself as a practical, scalable, and domain-oriented approach for ROP prediction,
contributing to the advancement of machine learning techniques in petroleum engineering.
